荒漠草原区土壤水分ELM-ESTARFM遥感反演模型构建与应用

王欢; 李瑞平; 王福强; 赵建伟; 苗存立; 籍晓婧

王欢,李瑞平,王福强,赵建伟,苗存立,籍晓婧.荒漠草原区土壤水分ELM-ESTARFM遥感反演模型构建与应用[J].干旱地区农业研究,2024,(3):236~244

荒漠草原区土壤水分ELM-ESTARFM遥感反演模型构建与应用

Construction and application of ELM-ESTARFM remote sensing inversion model for soil moisture in desert steppe area

DOI：10.7606/j.issn.1000-7601.2024.03.25

中文关键词: 土壤水分环境因子极限学习机随机森林

英文关键词:soil moisture environmental factors extreme learning machine random forest

基金项目:内蒙古自治区自然科学基金项目（2022MS05044）;国家自然科学基金项目（52269004）

作者	单位
王欢	内蒙古农业大学水利与土木建筑工程学院，内蒙古呼和浩特 010018
李瑞平	内蒙古农业大学水利与土木建筑工程学院，内蒙古呼和浩特 010018
王福强	内蒙古农业大学水利与土木建筑工程学院，内蒙古呼和浩特 010018
赵建伟	内蒙古农业大学水利与土木建筑工程学院，内蒙古呼和浩特 010018
苗存立	赤峰市克什克腾旗自然资源局，内蒙古赤峰 025350
籍晓婧	内蒙古自治区测绘地理信息中心，内蒙古呼和浩特 010018

摘要点击次数: 76

全文下载次数: 72

中文摘要:

土壤水分的高时空分辨率和高精度评估对干旱监测具有重要意义。为探究我国内蒙古荒漠草原区土壤水分遥感反演最优模型，基于Landsat和MODIS数据进行改进型自适应反射率时空融合（Enhanced Spatial and Temporal Adaptive Reflectance Fusion Model，ESTARFM），结合下垫面因子、地形因子、气象因子、植被因子等多要素环境因子，通过极限学习机（Extreme learning machine，ELM）和随机森林（Random forest，RF）两种方法构建土壤含水率反演模型，并与Landsat（未进行融合）构建的土壤含水率反演模型进行对比，最终筛选得到最优土壤含水率反演模型，并对研究区不同土地利用类型土壤含水率分布特征进行应用分析。结果表明：归一化植被指数是土壤含水率环境因子相关分析中最重要的预测因子（0～10、10～20、20～30 cm土壤深度处R²=0.85、0.82、0.79），其次为降水量（R²=0.73、0.68、0.71）、高程（R²=0.71、0.70、0.71）、水体指数（R²=0.69、0.69、0.68）、归一化盐分指数（R²=0.68、0.67、0.65）。与未进行时空融合所构建的模型相比，利用ESTARFM时空融合所构建的模型精度均有所提升，考虑ESTARFM时空融合时，ELM模型的R²、RMSE、MAE分别为0.89、6.58%、3.93%，RF模型的R²、RMSE、MAE分别为0.78、7.25%、4.95%;未考虑ESTARFM时空融合时，ELM模型的R²、RMSE、MAE分别为0.75、7.37%、5.24%，RF模型的R²、RMSE、MAE分别为0.71、7.48%、5.30%。表明ELM模型比RF模型的土壤含水率反演效果更好，且ELM-ESTARFM为土壤含水率反演最优模型。在此基础上，运用改进后的ELM-ESTARFM遥感反演模型监测了乌审旗全域土壤含水率，发现研究区北部和西北部的土壤含水率较高，南部地区的土壤含水率较低；对于不同土壤深度，土壤含水率由大到小依次为耕地、林地、草地、沙地，耕地区域0～10、10～20、20～30 cm土层含水率分别为18.92%、19.34%、21.84%，林地为11.80%、11.87%、12.40%，草地为10.97%、11.02%、12.22%，沙地为5.07%、5.35%、5.67%。

英文摘要:

The high spatiotemporal resolution and high\|precision assessment of soil moisture are of great significance for drought monitoring. To explore the optimal model of remote sensing inversion for soil moisture in the desert steppe region of Inner Mongolia, China, the Enhanced Spatial and Temporal Adaptive Reflectance Fusion Model (ESTARFM) was carried out based on Landsat and MODIS data. Combining with multi\|factor environmental factors including underlying surface factors, topographic factors, meteorological factors, and vegetation factors,the soil moisture content inversion model was constructed by extreme learning machine (ELM) and random forest (RF) methods. Comparing with the soil moisture content inversion model constructed by Landsat (without fusion), the optimal soil moisture content inversion model was selected. The soil moisture content distribution characteristics of different land use types in the study area were analyzed. The results showed that the normalized vegetation index was the most important predictor of soil moisture content and environmental factors (R²=0.85, 0.82, 0.79 at soil depth of 0~10, 10~20, 20~30 cm), followed by precipitation (R²=0.73, 0.68, 0.71), elevation (R²=0.71, 0.70, 0.71), water index (R²=0.69, 0.69, 0.68), and normalized salinity index (R²=0.68, 0.67、0.65). Compared with the model without spatiotemporal fusion, the accuracy of the model constructed by ESTARFM spatiotemporal fusion was improved, and the R², RMSE, and MAE of the ELM model were 0.89, 6.58%, and 3.93%, respectively, and the R², RMSE, and MAE of the RF model were 0.78, 7.25%, and 4.95%, respectively. The MAE was 0.75, 7.37%, and 5.24%, respectively, and the R², RMSE, and MAE of the RF model were 0.71, 7.48%, and 5.30%, respectively, indicating that the ELM model had a better inversion effect on soil moisture content than the RF model, and ELM-ESTARFM was the optimal model for soil moisture content inversion. On this basis, the improved ELM-ESTARFM remote sensing inversion model was used to monitor the soil moisture content in the whole area of Wushenqi, and it was found that the soil moisture content in the north and northwest of the study area was high, and the soil moisture content in the southern area was low. For different soil depths, the soil moisture content was 18.92%, 19.34%, and 21.84% in the cultivated land area at 0~10, 10~20, and 20~30 cm soil depths, 11.80%, 11.87%, 12.40% in woodland, 10.97%, 11.02% and 12.22% in grassland, and 5.07%, 5.35%, and 5.67% in sandy land, respectively.

查看全文查看/发表评论下载PDF阅读器